The document analyzes the KDD Cup 2009 competition, which involved predicting customer behavior using a large customer database from Orange, a French telecommunications company. The competition attracted over 450 participants from 46 countries. Ensemble methods such as decision trees were very effective at handling the large dataset that had many samples, attributes, mixed variable types, and missing values. The top-performing models were able to rapidly build predictive models and score new entries on the large, noisy, and unbalanced customer database.
Running head CASE STUDY QUESTIONS 1CASE STUDY QUESTIONS7.docxhealdkathaleen
Running head: CASE STUDY QUESTIONS 1
CASE STUDY QUESTIONS7
Case study questions
Name
Institution
Introduction
The CARE International is known to be a humanitarian organization which offers relief help to regions that are experiencing natural disasters like hurricanes and earthquakes. This paper will focus on answering the case study questions on diverse issues.
1. What were the main challenges encountered by CARE International before they created their warehouse prepositioning model
The challenges faced by CARE international are such as transportation issues, absence of agility, limited aptitude to gather supplies both locally and internationally (Gay, 2015). There is a challenge of the previous means of transportation utilized by vendors which is slow and also very unreliable as well as the cost of relief resources is very high.
2. How does the objective function relate to the organization's need to improve relief services to affected areas?
The objective function need to minimize the time used to respond to natural occurrences as well as transport reliefs to the affected area. The objective function is connected to the need of agency providing shipping relief services to the affected region at minimal response time.
3. Conduct online research and suggest at least three other applications or types of software that could handle the magnitude of variable and constraints CARE International used in their MIP model.
There are three software that can be incorporated to handle constraints Care International utilizes for Mixed-integer programming ideal which are “Mathematical Programming Language (AMPL), Gurobi, and Statistical Analysis System (SAS) (De & Mok, 2016).” These are tools that assist the organization in solving the linear and the quadratic optimization.
4. Elaborate on some benefits CARE International stands to gain from implementing their pre-positioning model on a large scale in future
Warehouse pre-positioning model has some advantages that it offers to CARE International such as they facilitate the delivery relief to the affected region, mostly in the time of the disaster occurrence. This model assist in providing relief in the affected region in relatively short period of time which means that the agency will be in a position to save more lives.
Warehouse pre-positioning model
Figure 1: Warehouse pre-positioning model
Application case
1. Describe the problem that a large company such as HP might face in offering many product lines and options.
One problem they might face is the issue of high cost for introducing new products, designing and manufacturing as well as the issue of supply chain complexity. The institution might face issues such as the unexpected increase in operational expenses, reduction in the forecasting accuracy and increase in the inventory costs.
2. Why is there a possible conflict between marketing and operations?
Yes. Marketing and operation department are the pillars of any organization. How ...
Running head CASE STUDY QUESTIONS 1CASE STUDY QUESTIONS7.docxhealdkathaleen
Running head: CASE STUDY QUESTIONS 1
CASE STUDY QUESTIONS7
Case study questions
Name
Institution
Introduction
The CARE International is known to be a humanitarian organization which offers relief help to regions that are experiencing natural disasters like hurricanes and earthquakes. This paper will focus on answering the case study questions on diverse issues.
1. What were the main challenges encountered by CARE International before they created their warehouse prepositioning model
The challenges faced by CARE international are such as transportation issues, absence of agility, limited aptitude to gather supplies both locally and internationally (Gay, 2015). There is a challenge of the previous means of transportation utilized by vendors which is slow and also very unreliable as well as the cost of relief resources is very high.
2. How does the objective function relate to the organization's need to improve relief services to affected areas?
The objective function need to minimize the time used to respond to natural occurrences as well as transport reliefs to the affected area. The objective function is connected to the need of agency providing shipping relief services to the affected region at minimal response time.
3. Conduct online research and suggest at least three other applications or types of software that could handle the magnitude of variable and constraints CARE International used in their MIP model.
There are three software that can be incorporated to handle constraints Care International utilizes for Mixed-integer programming ideal which are “Mathematical Programming Language (AMPL), Gurobi, and Statistical Analysis System (SAS) (De & Mok, 2016).” These are tools that assist the organization in solving the linear and the quadratic optimization.
4. Elaborate on some benefits CARE International stands to gain from implementing their pre-positioning model on a large scale in future
Warehouse pre-positioning model has some advantages that it offers to CARE International such as they facilitate the delivery relief to the affected region, mostly in the time of the disaster occurrence. This model assist in providing relief in the affected region in relatively short period of time which means that the agency will be in a position to save more lives.
Warehouse pre-positioning model
Figure 1: Warehouse pre-positioning model
Application case
1. Describe the problem that a large company such as HP might face in offering many product lines and options.
One problem they might face is the issue of high cost for introducing new products, designing and manufacturing as well as the issue of supply chain complexity. The institution might face issues such as the unexpected increase in operational expenses, reduction in the forecasting accuracy and increase in the inventory costs.
2. Why is there a possible conflict between marketing and operations?
Yes. Marketing and operation department are the pillars of any organization. How ...
Il ruolo chiave degli Advanced Analytics per la Supply ChainACTOR
Presentation of the speech held by Raffaele Maccioni (Co-Founder and CEO at ACT Operations Research) and Claudia Beldon (VP - Fashion & Luxury Industry at ACT Operations Research) at the recent "Logistica Efficiente" event titled "L'innovazione nella Supply Chain 2018".
Cloud computing business framework
Victor Chang, Leeds Beckett University
International conference on
“DATA, DIGITAL BUSINESS MODELS, CLOUD COMPUTING AND ORGANIZATIONAL DESIGN”
24-25 November 2014 ,
Université Paris –Sud
Churn in the Telecommunications Industryskewdlogix
Strategic Business Analysis Capstone Project Telecommunications Churn Management
Churn is a significant problem that costs telecommunications companies billions of dollars through lost revenue. Now that the market is more mature, the only way for a company to grow is to take their competitors customers. This issue
combined with the greater choice that consumers have gained means that any adverse touch point with a consumer can result in a lost customer.
Dear students get fully solved assignments by professionals
Send your semester & Specialization name to our mail id :
stuffstudy5@gmail.com
or
call us at : 098153-33456
Dear students get fully solved assignments
Send your semester & Specialization name to our mail id :
“ help.mbaassignments@gmail.com ”
or
Call us at : 08263069601
Bba501 & production and operations managementsmumbahelp
Dear students get fully solved assignments
Send your semester & Specialization name to our mail id :
help.mbaassignments@gmail.com
or
call us at : 08263069601
Dear students get fully solved assignments
Send your semester & Specialization name to our mail id :
“ help.mbaassignments@gmail.com ”
or
Call us at : 08263069601
(Prefer mailing. Call in emergency )
Visual and analytical mining of sales transaction data for production plannin...Gurdal Ertek
Recent developments in information technology paved the way for the collection of large amounts of data pertaining to various aspects of an enterprise. The greatest challenge faced in
processing these massive amounts of raw data gathered turns out to be the effective management of data with the ultimate purpose of deriving necessary and meaningful information
out of it. The following paper presents an attempt to illustrate the combination of visual and analytical data mining techniques for planning of marketing and production activities. The
primary phases of the proposed framework consist of filtering, clustering and comparison steps implemented using interactive pie charts, K-Means algorithm and parallel coordinate plots
respectively. A prototype decision support system is developed and a sample analysis session is conducted to demonstrate the applicability of the framework.
http://research.sabanciuniv.edu.
Customer churn classification using machine learning techniquesSindhujanDhayalan
Advanced data mining project on classifying customer churn by
using machine learning algorithms such as random forest,
C5.0, Decision tree, KNN, ANN, and SVM. CRISP-DM approach was followed for developing the project. Accuracy rate, Error rate, Precision, Recall, F1 and ROC curve was generated using R programming and the efficient model was found comparing these values.
Data Mining on Customer Churn ClassificationKaushik Rajan
Implemented multiple classifiers to classify if a customer will leave or stay with the company based on multiple independent variables.
Tools used:
> RStudio for Exploratory data analysis, Data Pre-processing and building the models
> Tableau and RStudio for Visualization
> LATEX for documentation
Machine learning models used:
> Random Forest
> C5.0
> Decision tree
> Neural Network
> K-Nearest Neighbour
> Naive Bayes
> Support Vector Machine
Methodology: CRISP-DM
Broadband is a relatively new technology, and its underlying science is still being developed. We have long understood the 'right' units in other engineering disciplines: mass, length, hardness, etc. What is the 'right' unit for supply and demand for broadband?
This presentation discusses the need for having the right metric. This means solving two problems: the 'abstraction' gap, and the 'inference' gap. ∆Q is the ideal metric because it fills both gaps.
Il ruolo chiave degli Advanced Analytics per la Supply ChainACTOR
Presentation of the speech held by Raffaele Maccioni (Co-Founder and CEO at ACT Operations Research) and Claudia Beldon (VP - Fashion & Luxury Industry at ACT Operations Research) at the recent "Logistica Efficiente" event titled "L'innovazione nella Supply Chain 2018".
Cloud computing business framework
Victor Chang, Leeds Beckett University
International conference on
“DATA, DIGITAL BUSINESS MODELS, CLOUD COMPUTING AND ORGANIZATIONAL DESIGN”
24-25 November 2014 ,
Université Paris –Sud
Churn in the Telecommunications Industryskewdlogix
Strategic Business Analysis Capstone Project Telecommunications Churn Management
Churn is a significant problem that costs telecommunications companies billions of dollars through lost revenue. Now that the market is more mature, the only way for a company to grow is to take their competitors customers. This issue
combined with the greater choice that consumers have gained means that any adverse touch point with a consumer can result in a lost customer.
Dear students get fully solved assignments by professionals
Send your semester & Specialization name to our mail id :
stuffstudy5@gmail.com
or
call us at : 098153-33456
Dear students get fully solved assignments
Send your semester & Specialization name to our mail id :
“ help.mbaassignments@gmail.com ”
or
Call us at : 08263069601
Bba501 & production and operations managementsmumbahelp
Dear students get fully solved assignments
Send your semester & Specialization name to our mail id :
help.mbaassignments@gmail.com
or
call us at : 08263069601
Dear students get fully solved assignments
Send your semester & Specialization name to our mail id :
“ help.mbaassignments@gmail.com ”
or
Call us at : 08263069601
(Prefer mailing. Call in emergency )
Visual and analytical mining of sales transaction data for production plannin...Gurdal Ertek
Recent developments in information technology paved the way for the collection of large amounts of data pertaining to various aspects of an enterprise. The greatest challenge faced in
processing these massive amounts of raw data gathered turns out to be the effective management of data with the ultimate purpose of deriving necessary and meaningful information
out of it. The following paper presents an attempt to illustrate the combination of visual and analytical data mining techniques for planning of marketing and production activities. The
primary phases of the proposed framework consist of filtering, clustering and comparison steps implemented using interactive pie charts, K-Means algorithm and parallel coordinate plots
respectively. A prototype decision support system is developed and a sample analysis session is conducted to demonstrate the applicability of the framework.
http://research.sabanciuniv.edu.
Customer churn classification using machine learning techniquesSindhujanDhayalan
Advanced data mining project on classifying customer churn by
using machine learning algorithms such as random forest,
C5.0, Decision tree, KNN, ANN, and SVM. CRISP-DM approach was followed for developing the project. Accuracy rate, Error rate, Precision, Recall, F1 and ROC curve was generated using R programming and the efficient model was found comparing these values.
Data Mining on Customer Churn ClassificationKaushik Rajan
Implemented multiple classifiers to classify if a customer will leave or stay with the company based on multiple independent variables.
Tools used:
> RStudio for Exploratory data analysis, Data Pre-processing and building the models
> Tableau and RStudio for Visualization
> LATEX for documentation
Machine learning models used:
> Random Forest
> C5.0
> Decision tree
> Neural Network
> K-Nearest Neighbour
> Naive Bayes
> Support Vector Machine
Methodology: CRISP-DM
Broadband is a relatively new technology, and its underlying science is still being developed. We have long understood the 'right' units in other engineering disciplines: mass, length, hardness, etc. What is the 'right' unit for supply and demand for broadband?
This presentation discusses the need for having the right metric. This means solving two problems: the 'abstraction' gap, and the 'inference' gap. ∆Q is the ideal metric because it fills both gaps.
Synthetic Fiber Construction in lab .pptxPavel ( NSTU)
Synthetic fiber production is a fascinating and complex field that blends chemistry, engineering, and environmental science. By understanding these aspects, students can gain a comprehensive view of synthetic fiber production, its impact on society and the environment, and the potential for future innovations. Synthetic fibers play a crucial role in modern society, impacting various aspects of daily life, industry, and the environment. ynthetic fibers are integral to modern life, offering a range of benefits from cost-effectiveness and versatility to innovative applications and performance characteristics. While they pose environmental challenges, ongoing research and development aim to create more sustainable and eco-friendly alternatives. Understanding the importance of synthetic fibers helps in appreciating their role in the economy, industry, and daily life, while also emphasizing the need for sustainable practices and innovation.
Palestine last event orientationfvgnh .pptxRaedMohamed3
An EFL lesson about the current events in Palestine. It is intended to be for intermediate students who wish to increase their listening skills through a short lesson in power point.
This is a presentation by Dada Robert in a Your Skill Boost masterclass organised by the Excellence Foundation for South Sudan (EFSS) on Saturday, the 25th and Sunday, the 26th of May 2024.
He discussed the concept of quality improvement, emphasizing its applicability to various aspects of life, including personal, project, and program improvements. He defined quality as doing the right thing at the right time in the right way to achieve the best possible results and discussed the concept of the "gap" between what we know and what we do, and how this gap represents the areas we need to improve. He explained the scientific approach to quality improvement, which involves systematic performance analysis, testing and learning, and implementing change ideas. He also highlighted the importance of client focus and a team approach to quality improvement.
Unit 8 - Information and Communication Technology (Paper I).pdfThiyagu K
This slides describes the basic concepts of ICT, basics of Email, Emerging Technology and Digital Initiatives in Education. This presentations aligns with the UGC Paper I syllabus.
Instructions for Submissions thorugh G- Classroom.pptxJheel Barad
This presentation provides a briefing on how to upload submissions and documents in Google Classroom. It was prepared as part of an orientation for new Sainik School in-service teacher trainees. As a training officer, my goal is to ensure that you are comfortable and proficient with this essential tool for managing assignments and fostering student engagement.
Model Attribute Check Company Auto PropertyCeline George
In Odoo, the multi-company feature allows you to manage multiple companies within a single Odoo database instance. Each company can have its own configurations while still sharing common resources such as products, customers, and suppliers.
How to Split Bills in the Odoo 17 POS ModuleCeline George
Bills have a main role in point of sale procedure. It will help to track sales, handling payments and giving receipts to customers. Bill splitting also has an important role in POS. For example, If some friends come together for dinner and if they want to divide the bill then it is possible by POS bill splitting. This slide will show how to split bills in odoo 17 POS.
The French Revolution, which began in 1789, was a period of radical social and political upheaval in France. It marked the decline of absolute monarchies, the rise of secular and democratic republics, and the eventual rise of Napoleon Bonaparte. This revolutionary period is crucial in understanding the transition from feudalism to modernity in Europe.
For more information, visit-www.vavaclasses.com
1. JMLR: Workshop and Conference Proceedings 7: 1-22
KDD cup 2009
Analysis of the KDD Cup 2009:
Fast Scoring on a Large Orange Customer Database
Isabelle Guyon
isabelle@clopinet.com
Clopinet, Berkeley, CA 94798, USA
Vincent Lemaire
vincent.lemaire@orange-ftgroup.com
Orange Labs, Lannion, 22300, France
Marc Boull´
e
marc.boulle@orange-ftgroup.com
Orange Labs, Lannion, 22300, France
Gideon Dror
gideon@mta.ac.il
Academic College of Tel-Aviv-Yaffo, Tel Aviv 61083, Israel
David Vogel
dvogel@dataminingsolutions.net
Data Mining Solutions, Orlando, Florida, USA
Editor: Neil Lawrence
Abstract
We organized the KDD cup 2009 around a marketing problem with the goal of identifying
data mining techniques capable of rapidly building predictive models and scoring new
entries on a large database. Customer Relationship Management (CRM) is a key element of
modern marketing strategies. The KDD Cup 2009 offered the opportunity to work on large
marketing databases from the French Telecom company Orange to predict the propensity
of customers to switch provider (churn), buy new products or services (appetency), or buy
upgrades or add-ons proposed to them to make the sale more profitable (up-selling). The
challenge started on March 10, 2009 and ended on May 11, 2009. This challenge attracted
over 450 participants from 46 countries. We attribute the popularity of the challenge to
several factors: (1) A generic problem relevant to the Industry (a classification problem),
but presenting a number of scientific and technical challenges of practical interest including:
a large number of training examples (50,000) with a large number of missing values (about
60%) and a large number of features (15,000), unbalanced class proportions (fewer than
10% of the examples of the positive class), noisy data, presence of categorical variables
with many different values. (2) Prizes (Orange offered 10,000 Euros in prizes). (3) A
well designed protocol and web site (we benefitted from past experience). (4) An effective
advertising campaign using mailings and a teleconference to answer potential participants
questions. The results of the challenge were discussed at the KDD conference (June 28,
2009). The principal conclusions are that ensemble methods are very effective and that
ensemble of decision trees offer off-the-shelf solutions to problems with large numbers of
samples and attributes, mixed types of variables, and lots of missing values. The data and
the platform of the challenge remain available for research and educational purposes at
http://www.kddcup-orange.com/.
Keywords: challenge, classification, customer management, fast scoring
c 2009 Isabelle Guyon, Vincent Lemaire, Marc Boull´, Gideon Dror, David Vogel.
e
2. ´
Guyon, Lemaire, Boulle, Dror, Vogel
1. Introduction
Customer Relationship Management (CRM) is a key element of modern marketing strategies. The KDD Cup 2009 offered the opportunity to work on large marketing databases
from the French Telecom company Orange to predict the propensity of customers to switch
provider (churn), buy new products or services (appetency), or buy upgrades or add-ons
proposed to them to make the sale more profitable (up-selling).
The most practical way to build knowledge on customers in a CRM system is to produce
scores. A score (the output of a model) is an evaluation for all target variables to explain
(i.e., churn, appetency or up-selling). Tools producing scores provide quantifiable information on a given population. The score is computed using customer records represented by
a number of variables or features. Scores are then used by the information system (IS),
for example, to personalize the customer relationship. The rapid and robust detection of
the most predictive variables can be a key factor in a marketing application. An industrial
customer analysis platform developed at Orange Labs, capable of building predictive models for datasets having a very large number of input variables (thousands) and instances
(hundreds of thousands), is currently in use by Orange marketing. A key requirement is the
complete automation of the whole process. The system extracts a large number of features
from a relational database, selects a subset of informative variables and instances, and efficiently builds in a few hours an accurate classifier. When the models are deployed, the
platform exploits sophisticated indexing structures and parallelization in order to compute
the scores of millions of customers, using the best representation.
The challenge was to beat the in-house system developed by Orange Labs. It was
an opportunity for participants to prove that they could handle a very large database,
including heterogeneous noisy data (numerical and categorical variables), and unbalanced
class distributions. Time efficiency is often a crucial point. Therefore part of the competition
was time-constrained to test the ability of the participants to deliver solutions quickly. The
fast track of the challenge lasted five days only. To encourage participation, the slow track
of the challenge allowed participants to continue working on the problem for an additional
month. A smaller database was also provided to allow participants with limited computer
resources to enter the challenge.
2. Background and motivations
This challenge uses important marketing problems to benchmark classification methods in
a setting, which is typical of large-scale industrial applications. A large database was made
available by the French Telecom company, Orange with tens of thousands of examples and
variables. This dataset is unusual in that it has a large number of variables making the
problem particularly challenging to many state-of-the-art machine learning algorithms. The
challenge participants were provided with masked customer records and their goal was to
predict whether a customer will switch provider (churn), buy the main service (appetency)
and/or buy additional extras (up-selling), hence solving three binary classification problems.
Churn is the propensity of customers to switch between service providers, appetency is the
propensity of customers to buy a service, and up-selling is the success in selling additional
good or services to make a sale more profitable. Although the technical difficulty of scaling
2
3. Analysis of the KDD Cup 2009
up existing algorithms is the main emphasis of the challenge, the dataset proposed offers
a variety of other difficulties: heterogeneous data (numerical and categorical variables),
noisy data, unbalanced distributions of predictive variables, sparse target values (only 1 to
7 percent of the examples examples belong to the positive class) and many missing values.
3. Evaluation
There is value in a CRM system to evaluate the propensity of customers to buy. Therefore,
tools producing scores are more usable that tools producing binary classification results. The
participants were asked to provide a score (a discriminant value or a posterior probability
P (Y = 1|X)), and they were judged by the area under the ROC curve (AUC). The AUC
is the area under the curve plotting sensitivity vs. (1− specificity) when the threshold
θ is varied (or equivalently the area under the curve plotting sensitivity vs. specificity).
We call “sensitivity” the error rate of the positive class and “specificity” the error rate
of the negative class. The AUC is a standard metric in classification. There are several
ways of estimating error bars for the AUC. We used a simple heuristic, which gives us
approximate error bars, and is fast and easy to implement: we find on the AUC curve the
point corresponding to the largest balanced accuracy BAC = 0.5 (sensitivity + specificity).
We then estimate the standard deviation of the BAC as:
σ=
1
2
p+ (1 − p+ ) p− (1 − p− )
+
,
m+
m−
(1)
where m+ is the number of examples of the positive class, m− is the number of examples
of the negative class, and p+ and p− are the probabilities of error on examples of the
positive and negative class, approximated by their empirical estimates, the sensitivity and
the specificity (Guyon et al., 2006).
The fraction of positive/negative examples posed a challenge to the participants, yet it
was sufficient to ensure robust prediction performances (as verified in the beta tests). The
database consisted of 100,000 instances, split randomly into equally sized train and test
sets:
• Churn problem: 7.3% positive instances (3672/50000 on train).
• Appetency problem: 1.8% positive instances (890/50000 on train).
• Up-selling problem: 7.4% positive instances (3682/50000 on train).
On-line feed-back on AUC performance was provided to the participants who made
correctly formatted submissions, using only 10% of the test set. There was no limitation
on the number of submissions, but only the last submission on the test set (for each task)
was taken into account for the final ranking.
The score used for the final ranking was the average of the scores on the three tasks
(churn, appetency, and up-selling).
4. Data
Orange (the French Telecom company) made available a large dataset of customer data,
each consisting of:
3
4. ´
Guyon, Lemaire, Boulle, Dror, Vogel
• Training : 50,000 instances including 15,000 inputs variables, and the target value.
• Test : 50,000 instances including 15,000 inputs variables.
There were three binary target variables (corresponding to churn, appetency, and upselling). The distribution within the training and test examples was the same (no violation
of the i.i.d. assumption - independently and identically distributed). To encourage participation, an easier task was also built from a reshuffled version of the datasets with only
230 variables. Hence, two versions were made available (“small” with 230 variables, and
“large” with 15,000 variables). The participants could enter results on either or both versions, which corresponded to the same data entries, the 230 variables of the small version
being just a subset of the 15,000 variables of the large version. Both training and test data
were available from the start of the challenge, without the true target labels. For practice
purposes, “toy” training labels were available together with the training data from the onset of the challenge in the fast track. The results on toy targets did not count for the final
evaluation. The real training labels of the tasks “churn”, “appetency”, and “up-selling”,
were later made available for download, half-way through the challenge.
The database of the large challenge was provided in several chunks to be downloaded
more easily and we provided several data mirrors to avoid data download congestion.
The data were made publicly available through the website of the challenge http://www.
kddcup-orange.com/, with no restriction of confidentiality. They are still available to
download for benchmark purpose. To protect the privacy of the customers whose records
were used, the data were anonymized by replacing actual text or labels by meaningless
codes and not revealing the meaning of the variables.
Extraction and preparation of the challenge data:
The Orange in-house customer analysis platform is devoted to industrializing the data mining process for marketing purpose. Its fully automated data processing machinery includes:
data preparation, model building, and model deployment. The data preparation module
was isolated and used to format data for the purpose of the challenge and facilitate the task
of the participants. Orange customer data are initially available in a relational datamart
under a star schema. The platform uses a feature construction language, dedicated to the
marketing domain, to build tens of thousands of features to create a rich data representation
space.
For the challenge, a datamart of about one million of customers was used, with about
ten tables and hundreds of fields. The first step was to resample the dataset, to obtain
100,000 instances with less unbalanced target distributions. For practical reasons (the
challenge participants had to download the data), the same data sample was used for the
three marketing tasks. In a second step, the feature construction language was used to
generate 20,000 features and obtain a tabular representation. After discarding constant
features and removing customer identifiers, we narrowed down the feature set to 15,000
variables (including 260 categorical variables). In a third step, for privacy reasons, data was
anonymized, discarding variables names, randomizing the order of the variables, multiplying
each continuous variable by a random factor and recoding categorical variable with randomly
generated category names. Finally, the data sample was split randomly into equally sized
4
5. Analysis of the KDD Cup 2009
train and test sets. A random subset of 10% of the test set was designated to provide
immediate performance feed-back.
5. Beta tests
The website of the challenge http://www.kddcup-orange.com/ was thoroughly tested by
the KDD cup chairs and volunteers. The datasets were downloaded and checked. Baseline
methods were tried to verify the feasibility of the task. A Matlab R version of the data was
made available and sample code were provided to format the results. A sample submission
of random results was given as example and submitted to the website. The results of the
Na¨ Bayes method were also uploaded to the website to provide baseline results.
ıve
Toy problem:
The Toy problem on the LARGE dataset consisted of one single predictive continuous
variable (V5963) uniformly distributed on the interval [0, 2.0]. The target value was obtained
by thresholding V5963 at 1.6 and adding 20% noise. Hence for 80% of the instances, lying
in interval [0, 1.6], the fraction of positive examples is 20%; for the remaining 20% lying
in interval ]1.6, 2.0], the fraction of positive examples is 80%. The expected value of the
AUC (called “true AUC”) can easily be computed1 . Its value is approximately 0.7206.
Because of the variance in the sampling process, the AUC effectively computed using the
optimal decision rule (called “optimal AUC”) is 0.7196 for the training set and a 0.7230
for the test set. Interestingly, as shown in Figure 1, the optimal solution was outperformed
by many participants, up to 0.7263. This illustrates the problem of multiple testing and
shows how the best test performance overestimates both the expected value of the AUC
and the performance of the optimal decision rule, increasingly with the number of challenge
submissions.
Basic Na¨ Bayes classifier:
ıve
The basic Na¨ Bayes classifier (see e.g., Mitchell, 1997) makes simple independence asıve
sumptions between features and votes among features with a voting score capturing the
correlation of the feature to the target. No feature selection is performed and there are no
hyper-parameters to adjust.
For the LARGE dataset, the overall score of the basic Na¨ Bayes classifier is 0.6711,
ıve
with the following results on the test set:
• Churn problem : AUC = 0.6468;
• Appetency problem : AUC = 0.6453;
1. If we call T the total number of examples, the (expected value of) the total number of examples of the
positive class P is the sum of the number of positive examples in the first and the second intervals,
i.e., P = (0.2 × 0.8 + 0.8 × 0.2) T = 0.32 T . Similarly, the total number of negative examples is
N = (0.8 × 0.8 + 0.2 × 0.2) T = 0.68 T . If we use the optimal decision rule (a threshold on V5963 at
1.6) the number of true positive examples is the sum of the number of true positive examples in the
two intervals, i.e., T P = 0 + (0.2 × 0.8) T = 0.16 T . Similarly, the number of true negative examples is
T N = (0.8 × 0.8) T = 0.64 T . Hence, the true positive rate is T P R = T P/P = 0.16/0.32 = 0.5 and the
true negative rate is T N R = T N/N = 0.64/0.68 0.9412. The balanced accuracy (or the AUC because
BAC = AU C in this case) is therefore: BAC = 0.5 (T P R + T N R) = 0.5 (0.5 + 0.9412) = 0.7206.
5
6. ´
Guyon, Lemaire, Boulle, Dror, Vogel
Test AUC
0.73
Test
True
Optimal
0.72
0.71
10/3/09
Date
17/3/09
24/3/09
31/3/09
7/4/09
14/4/09
Figure 1: Toy problem test results.
• Up-selling problem : AUC=0.7211;
As per the rules of the challenge, the participants had to outperform the basic Na¨
ıve
Bayes classifier to qualify for prizes.
Orange in-house classifier:
The Orange in-house classifier is an extension of the Na¨ Bayes classifier, called “Selective
ıve
Na¨ Bayes classifier” (Boull´, 2007). It includes an optimized preprocessing, variable
ıve
e
selection, and model averaging. It significantly outperforms the basic Na¨ Bayes classifier
ıve
performance, which was provided to the participants as baseline, and it is computationally
efficient: The results were obtained after 3 hours using a standard laptop, considering the
three tasks as three different problems. The models were obtained by applying the training
process Khiops R only once since the system has no hyper-parameter to adjust. The results
of the in-house system were not revealed until the end of the challenge. An implementation
of the method is available as shareware from http://www.khiops.com; some participants
downloaded it and used it.
The requirements placed on the in-house system are to obtain a high classification
accuracy, under the following constraints:
• Fully automatic: absolutely no hyper-parameter setting, since hundred of models need
to be trained each month.
• Fast to train: the three challenge marketing problems were trained in less than 3 hours
on a mono-processor laptop with 2 Go RAM.
• Efficient after deployment: models need to process rapidly up to ten million instances.
• Interpretable: selected predictive variables must provide insight.
6
7. Analysis of the KDD Cup 2009
However, for the challenge, the participants were not placed under all these constraints
for practical reasons: it would have been both too constraining for the participants and too
difficult to enforce for the organizers. The challenge focused on maximizing accuracy under
time constraints.
For the LARGE dataset, the overall score of the Orange in-house classifier is 0.8311,
with the following results on the test dataset:
• Churn problem : AUC = 0.7435;
• Appetency problem : AUC = 0.8522;
• Up-selling problem : AUC=0.8975;
The challenge was to beat these results, but the minimum requirement to win prizes
was only to outperform the basic Na¨ Bayes classifier.
ıve
6. Challenge schedule and protocol
The key elements of our design were:
• To make available the training and test data three weeks before the start of the “fast
challenge” to allow participants to download the large volume of data, read it and
preprocess it without the training labels.
• To make available “toy” training labels during that period so participants could finalize their methodology and practice using the on-line submission system.
• To put participants under time pressure once the training labels were released (produce
results in five days) to test their ability to produce results in a timely manner.
• To continue the challenge beyond this first milestone for another month (slow challenge) to give the opportunity to participants with less computational resources to
enter the challenge.
• To provide a down-sized version of the dataset for the slow challenge providing an
opportunity for participants with yet less computational resources to enter the challenge.
• To provide large prizes to encourage participation (10,000 Euros donated by Orange),
without any strings attached (no legal constraint or commitment to release code or
methods to download data or participate).
The competition rules are summarized below are inspired from previous challenges we
organized (Clopinet):
1. Conditions of participation: Anybody who complied with the rules of the challenge
(KDD cup 2009) was welcome to participate. Only the organizers listed on the Credits
page were excluded from participating. The participants were not required to attend
the KDD cup 2009 workshop and the workshop was open to anyone.
7
8. ´
Guyon, Lemaire, Boulle, Dror, Vogel
2. Anonymity: All entrants had to identify themselves by registering on the KDD cup
2009 website. However, they could elect to remain anonymous during the development
period. To be eligible for prizes the had to publicly reveal their identity. Teams had
to declare a team leader and register their members. No individual could be part of
two or more teams.
3. Data: The datasets were available for download from the Dataset page to registered
participants. The data were available in several archives to facilitate downloading.
4. Challenge duration and tracks: The challenge started March 10, 2009 and ended
May 11, 2009. There were two challenge tracks:
• FAST (large) challenge: Results submitted on the LARGE dataset within five
days of the release of the real training labels counted towards the fast challenge.
• SLOW challenge: Results on the small dataset and results on the large dataset
not qualifying for the fast challenge, submitted before the KDD cup 2009 deadline
May 11, 2009, counted toward the SLOW challenge.
If more than one submission was made in either track and with either dataset, the last
submission before the track deadline was taken into account to determine the ranking
of participants and attribute the prizes.
5. On-line feed-back: During the challenge, the training set performances were available on the Result page as well as partial information on test set performances: The
test set performances on the “toy problem” and performances on a fixed 10% subset of the test examples for the real tasks (churn, appetency and up-selling). After
the challenge was over, the performances on the whole test set were calculated and
substituted in the result tables.
6. Submission method: The method of submission was via the form on the Submission
page, following a designated format. Results on the “toy problem” did not count
as part of the competition. Multiple submissions were allowed, but limited to 5
submissions per day to avoid congestion. For the final entry in the slow track, the
participants could submit results on either (or both) small and large datasets in the
same archive.
7. Evaluation and ranking: For each entrant, only the last valid entry, as defined
in the Instructions counted towards determining the winner in each track (fast and
slow). We limited each participating person to a single final entry in each track. Valid
entries had to include results on all three real tasks. Prizes could be attributed only
to entries performing better than the baseline method (Na¨ Bayes). The results of
ıve
the baseline method were provided to the participants.
8. Reproducibility: Participation was not conditioned on delivering code nor publishing methods. However, we asked the participants to voluntarily fill out a fact sheet
about their methods and contribute papers to the proceedings.
8
9. Analysis of the KDD Cup 2009
Figure 2: KDD Cup Participation by year (number of teams).
The full rules are available from the website of the challenge http://www.kddcup-orange.
com/. The rules were designed to attract a large number of participants and were successful
in that respect: Many participants did not participate in the fast challenge on the large
dataset, but entered in the slow track, either on the small or the large dataset (or both).
There was one minor design mistake: the small dataset was derived from the same data as
the large one and, despite our efforts to disguise the identity of the features, it was possible for some entrants to match the features and entries in the small and large dataset.
This provided a small advantage, in the slow track only, to the teams who did that data
“unscrambling”: they could get feed-back on 20% of the data rather than 10%.
The schedule of the challenge was as follows (Dates in 2009):
• March 10 - Start of the FAST large challenge. Data tables without target values were
made available for the large dataset. Toy training target values were made available for
practice purpose. Objective: participants can download data, ask questions, finalize
their methodology, try the submission process.
• April 6 - Training target values were made available for the large dataset for the real
problems (churn, appetency, and upselling). Feed-back: results on 10% of the test set
available on-line when submissions are made.
• April 10 - Deadline for the FAST large challenge. Submissions had to be received
before midnight, time zone of the challenge web server.
• April 11 - Data tables and training target values were made available for the small
dataset. The challenge continued for the large dataset in the slow track.
• May 11 - Deadline for the SLOW challenge (small and large datasets). Submissions
had to be be received before midnight, time zone of the challenge web server.
7. Results
The 2009 KDD Cup attracted 1299 teams from 46 different countries. From those teams,
7865 valid entries were submitted by 453 different teams. The participation was more than
9
10. ´
Guyon, Lemaire, Boulle, Dror, Vogel
Prize
1
2
3
1
2
3
Table 1: Winning entries.
Country
Fast
Rank
IBM Research
USA
1
ID Analytics
USA
2
David Slate & Peter Frey
USA
3
University of Melbourne
Australia
27
Financial Engineering Group, Inc.
Japan
4
National Taiwan University
Taiwan
20
Team
track
Score
0.8493
0.8448
0.8443
0.8250
0.8443
0.8332
Slow
Rank
1
3
8
2
4
5
track
Score
0.8521
0.8479
0.8443
0.8484
0.8477
0.8461
three times greater than any KDD Cup in the past. Figure 2 represents the KDD Cup
participation by year. A large participation was a key element to validate the results and
for Orange to have a ranking of its in-house system; the challenge was very successful in
that respect.
7.1 Winners
The overall winner is the IBM Research team (IBM Research, 2009) who ranked first in
both tracks. Six prizes were donated by Orange to top ranking participants in the fast and
the slow tracks (see Table 1). As per the rules of the challenge, the same team could not
earn two prizes. If the ranking of a team entitled it to two prizes, it received the best of
the two and the next best ranking team received the other prize.
All the winning teams scored best on the large dataset (and most participants obtained
better results on the large dataset then on the small dataset). IBM Research, ID Analytics,
and National Taiwan University (NTU) “unscrambled” the small dataset. This may have
provided an advantage only to NTU since “unscrambling” affected only the slow track and
the two other teams won prizes in the fast track. We briefly comment on the results of the
winners.
Fast track:
• IBM Research: The winning entry (IBM Research, 2009) consisted of an ensemble
of a wide variety of classifiers, following (Caruana and Niculescu-Mizil, 2004; Caruana
et al., 2006). Effort was put into coding (most frequent values coded with binary
features, missing values replaced by mean, extra features constructed, etc.)
• ID Analytics, Inc.: One of the only teams to use a wrapper feature selection strategy, following a filter (Xie et al., 2009). The classifier was built from the commercial
TreeNet software by Salford Systems: an additive boosting decision tree technology.
Bagging was also used to gain additional robustness.
• David Slate & Peter Frey (Old dogs with new tricks): After a simple preprocessing (consisting in grouping of modalities or discretizing) and filter feature selection, this team used ensembles of decision trees, similar to Random Forests (Breiman,
2001).
10
11. Analysis of the KDD Cup 2009
Slow track:
• University of Melbourne: This team used for feature selection a cross-validation
method targeting the AUC and, for classification, boosting with classification trees
and shrinkage, using a Bernoulli loss (Miller et al., 2009).
• Financial Engineering Group, Inc.: Few details were released by the team about
their methods. They used grouping of modalities and a filter feature selection method
using the AIC criterion (Akaike, 1973). Classification was based on gradient treeclassifier boosting (Friedman, 2000).
• National Taiwan University: The team averaged the performances of three classifiers (Lo et al., 2009): (1) The solution of the joint multiclass problem with an
L1-regularized maximum entropy model. (2) AdaBoost with tree-based weak learners (Freund and Schapire, 1996). (3) Selective Na¨ Bayes (Boull´, 2007), which is
ıve
e
the in-house classifier of Orange (see Section 5).
7.2 Performance statistics
We now turn to the statistical analysis of the results of the participants. The main statistics
are summarized in Table 2.
In the figures of this section, we use the following color code:
1. Black: Submissions received.
2. Blue: Overall best submissions received. Referred to as T estAU C ∗∗ .
3. Red: Baseline result, obtained with the basic Na¨ Bayes classifier or NB, provided
ıve
by the organizers (see Section 5). The organizers consider that this result is easy to
improve. They imposed that the participants would outperform this result to win
prizes to avoid that a random submission would win a prize.
4. Green: Orange system result, obtained by the in-house Orange system with the
Selective Na¨ Bayes classifier or SNB (see Section 5).
ıve
Progress in performance
Figure 3.a presents the results of the first day. A good result, better than the baseline
result, is obtained after one hour and the in-house system is slightly outperformed after
seven hours. The improvement during the first day of the competition, after the first 7
hours, is small: from 0.8347 to 0.8385.
Figure 3.b presents the results over the first 5 days (FAST challenge). The performance
progresses from 0.8385 to 0.8493. The rush of submissions before the deadline is clearly
observable. Considering only the submission with Test AUC > 0.5 in the first 5 days, 30% of
the submissions had worse results than the baseline (basic Na¨ Bayes) and 91% had worse
ıve
results than the in-house system (AUC=0.8311). Only and 9% of the submissions had
better results than the in-house system.
11
12. ´
Guyon, Lemaire, Boulle, Dror, Vogel
Table 2: Best results and baselines. The first four lines show the best score T AU C ∗
(averaged over the three tasks), over increasing periods of time [0 : t]. For comparison we give the results of the basic Na¨ Bayes classifier (NB) and the in-house
ıve
Orange system (SNB). The best overall performance is T AU C ∗∗ = T AU C ∗ (36d).
The relative performance difference δ ∗ = (T AU C ∗∗ − T AU C)/T AU C ∗∗ is given in
parenthesis (in percentage). The two last lines represent the relative performance
difference δ = (T AU C ∗ (t) − T AU C)/T AU C ∗ (t) for the two reference results.
T AU C (δ ∗ %)
Churn
Appetency
Up-selling
Average
δ NB %
δ SN B %
T AU C ∗ 12h
0.7467 (2.40)
0.8661 (2.17)
0.9011 (0.89)
0.8380 (1.65)
19.92
0.82
T AU C ∗ 24h
0.7467 (2.40)
0.8714 (1.57)
0.9011 (0.89)
0.8385 (1.60)
19.96
0.88
T AU C ∗ 5d
0.7611 (0.52)
0.8830 (0.26)
0.9057 (0.38)
0.8493 (0.33)
20.98
2.14
T AU C ∗∗
0.7651 (0)
0.8853 (0)
0.9092 (0)
0.8521 (0)
21.24
2.46
T AU C N B
0.6468 (15.46)
0.6453 (27.11)
0.7211 (20.69)
0.6711 (21.24)
-
T AU C SN B
0.7435 (2.82)
0.8522 (3.74)
0.8975 (1.29)
0.8311 (2.46)
-
Figures 3.a and 3.b show that good results were obtained already the first day and only
small improvements were made later. These results, and an examination of the fact sheets
of the challenge filled out by the participants, reveal that:
• There are available methods, which can process fast large databases using today’s
available hardware in both academia and industry.
• Several teams were capable of adapting their methods to meet the requirements of
the challenge and reach quickly good performances, yet the bulk of the participants
did not.
• The protocol of the challenge was well designed: the month given to download the
data and play with the submission protocol (using the toy problem) allowed us to
monitor progress in performance, not the time required to get ready for the challenge.
This last point was important for Orange to assess the time taken for generating stateof-the-art models, since speed of model generation is a key requirement in such applications.
The fact that performances do not significantly improve after a few hours is further confirmed in Figure 3.c: very small improvements (from 0.8493 to 0.8521) were made after the
5th day (SLOW challenge)2 .
Rapidity of model building
Figure 4.d gives a comparison between the submissions received and the best overall result
over increasing periods of time: 12 hours, one day, 5 days, and 36 days. We compute the
relative performance difference
δ ∗ = (T estAU C ∗∗ − T estAU C)/T estAU C ∗∗ ,
(2)
2. This improvement may be partly attributed to “unscrambling”; unscrambling was not possible during
the fast track of the challenge (first 5 days).
12
13. Analysis of the KDD Cup 2009
0.8
0.75
0.75
Test AUC
0.85
0.8
Test AUC
0.85
0.7
0.65
0.7
0.65
0.6
0.6
0.55
0.55
0.5
0
1
2
3
4
5
6
7
8
0.5
0
9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
24
Hour
48
72
96
120
Hour
(a) Test AUC - hour ∈ [0:24]
(b) Test AUC - day ∈ [1:5]
0.85
0.8
Test AUC
0.75
0.7
0.65
0.6
0.55
0.5
0
2
4
6
8
10 12 14 16 18 20 22 24 26 28 30 32 34 36
Day
(c) Test AUC - day ∈ [0:36]
Figure 3: Participant results over time. “Test AUC” is the average AUC on the test
set for the three problems. Each point represents an entry. The horizontal bars
represent: basic Na¨ Bayes, selective Na¨ Bayes, and best participant entry.
ıve
ıve
where T estAU C ∗∗ is the best overall result. The values of δ ∗ for the best performing
classifier in each interval and for the reference results are found in Table 2. The following
observations can be made:
• there is a wide spread of results;
• the median result improves significantly over time, showing that it is worth continuing
the challenge to give to participants an opportunity of learning how to solve the
problem (the median beats the baseline on all tasks after 5 days but keeps improving);
• but the best results do not improve a lot after the first day;
13
14. ´
Guyon, Lemaire, Boulle, Dror, Vogel
• and the distribution after 5 days is not very different from that after 36 days.
Table 2 reveals that, at the end of the challenge, for the average score, the relative
performance difference between the baseline model (basic Na¨ Bayes) and the best model
ıve
is over 20%, but only 2.46% for SNB. For the best ranking classifier, only 0.33% was gained
between the fifth day (FAST challenge) and the last day of the challenge (SLOW challenge).
After just one day, the best ranking classifier was only 1.60% away from the best result.
The in-house system (selective Na¨ Bayes) has a result less than 1% worse than the best
ıve
model after one day (δ = (1 − 0.8311/0.8385) = 0.88%).
We conclude that the participants did very well in building models fast. Building competitive models is one day is definitely doable and the Orange in-house system is competitive,
although it was rapidly beaten by the participants.
Individual task difficulty
To assess the relative difficulty of the three tasks, we plotted the relative performance
difference δ ∗ (Equation 2) for increasing periods of time, see Figure4.[a-c].
The churn task seems to be the most difficult one, if we consider that the performance
at day one, 0.7467, only increases to 0.7651 by the end of the challenge (see Table 2
for other intermediate results). Figure 4.a shows that the median performance after
one day is significantly worse than the baseline (Na¨ Bayes), whereas for the other
ıve
tasks the median was already beating the baseline after one day.
The appetency task is of intermediate difficulty. Its day one performance of 0.8714 increases to 0.8853 by the end of the challenge. Figure 4.b shows that, from day one,
the median performance beats the baseline method (which performs relatively poorly
on this task).
The up-selling task is the easiest one: the day one performance 0.9011, already very
high, improves to 0.9092 (less that 1% relative difference). Figure 4.c shows that, by
the end of the challenge, the median performance gets close to the best performance.
Correlation between T estAU C and V alidAU C:
The correlation between the results on the test set (100% of the test set), T estAU C, and the
results on the validation set (10% of the test set used to give a feed back to the competitors),
V alidAU C, is really good. This correlation (when keeping only AUC test results > 0.5
considering that test AUC < 0.5 are error submissions) is of 0.9960 ± 0.0005 (95% confidence
interval) for the first 5 days and of is of 0.9959 ± 0.0003 for the 36 days. These values indicate
that (i) the validation set was a good indicator for the online feedback; (ii) the competitors
have not overfitted the validation set. The analysis of the correlation indicator task by task
gives the same information, on the entire challenge (36 days) the correlation coefficient is
for the Churn task: 0.9860 ±0.001; for the Appetency task: 0.9875 ±0.0008 and for the
Up-selling task: 0.9974 ±0.0002.
Several participants studied the performance estimation variance by splitting the training data multiple times into 90% for training and 10% for validation. The variance in
14
15. Analysis of the KDD Cup 2009
80
60
60
Delta* (%)
100
80
Delta* (%)
100
40
20
0
40
20
12h
24h
5d
0
36d
12h
(a) Churn
5d
36d
(b) Appetency
100
80
80
60
60
Delta* (%)
100
Delta* (%)
24h
40
20
0
40
20
12h
24h
5d
0
36d
(c) Up-selling
12h
24h
5d
36d
(d) Average
Figure 4: Performance improvement over time. Delta∗ represents the relative difference in Test AUC compared to the overall best result T estAU C ∗∗ : Delta∗ =
(T estAU C ∗∗ − T estAU C)/T estAU C ∗∗ . On each box, the central mark is the
median, the edges of the box are the 25th and 75th percentiles, the whiskers extend to the most extreme data points not considered to be outliers; the outliers
are plotted individually as crosses.
the results that they obtained led them to use cross-validation to perform model selection
rather than relying on the 10% feed-back. Cross-validation was used by all the top ranking
participants. This may explain why the participants did not overfit the validation set.
We also asked the participants to return training set prediction results, hoping that
we could do an analysis of overfitting by comparing training set and test set performances.
However, because the training set results did not affect the ranking score, some participants
did not return real prediction performances using their classifier, but returned either random
results or the target labels. However, if we exclude extreme performances (random or
perfect), we can observe that (i) a fraction of the models performing well on test data
15
16. ´
Guyon, Lemaire, Boulle, Dror, Vogel
have a good correlation between training and test performances; (ii) there is a group of
models performing well on test data and having an AUC on training examples significantly
larger. Large margin models like SVMs (Boser et al., 1992) or boosting models (Freund
and Schapire, 1996) behave in this way. Among the models performing poorly on test data,
some clearly overfitted (had a large difference between training and test results).
7.3 Methods employed
We analyzed the information provided by the participants in the fact sheets. In Figures 5,6,7, and 8, we show histograms of the algorithms employed for preprocessing, feature
selection, classifier, and model selection. We briefly comment on these statistics:
• Preprocessing: Few participants did not use any preprocessing. A large fraction
of the participants replaced missing values by the mean or the median or a fixed
value. Some added an additional feature coding for the presence of a missing value.
This allows linear classifiers to automatically compute the missing value by selecting
an appropriate weight. Decision tree users did not replace missing values. Rather,
they relied on the usage of “surrogate variables”: at each split in a dichotomous tree,
if a variable has a missing value, it may be replaced by an alternative “surrogate”
variable. Discretization was the second most used preprocessing. Its usefulness for this
particular dataset is justified by the non-normality of the distribution of the variables
and the existence of extreme values. The simple bining used by the winners of the slow
track proved to be efficient. For categorical variables, grouping of under-represented
categories proved to be useful to avoid overfitting. The winners of the fast and the
slow track used similar strategies consisting in retaining the most populated categories
and coarsely grouping the others in an unsupervised way. Simple normalizations were
also used (like dividing by the mean). Principal Component Analysis (PCA) was
seldom used and reported not to bring performance improvements.
• Feature selection: Feature ranking and other filter methods were the most widely
used feature selection methods. Most participants reported that wrapper methods
overfitted the data. The winners of the slow track method used a simple technique
based on cross-validation classification performance of single variables.
• Classification algorithm: Ensembles of decision trees were the most widely used
classification method in this challenge. They proved to be particularly well adapted
to the nature of the problem: large number of examples, mixed variable types, and
lots of missing values. The second most widely used method was linear classifiers,
and more particularly logistic regression (see e.g., Hastie et al., 2000). Third came
non-linear kernel methods (e.g., Support Vector Machines, Boser et al. 1992). They
suffered from higher computational requirements, so most participants gave up early
on them and rather introduced non-linearities by building extra features.
• Model selection: The majority of the participants reported having used to some
extent the on-line performance feed-back on 10% of the test set for model selection.
However, the winners all declared that they quickly realized that due to variance in
the data, this method was unreliable. Cross-validation (ten-fold or five-fold) has been
16
17. Analysis of the KDD Cup 2009
PREPROCESSING (overall usage=95%)
Replacement of the missing values
Discretization
Normalizations
Grouping modalities
Other prepro
Principal Component Analysis
0
20
40
Percent of participants
60
80
Figure 5: Preprocessing methods.
FEATURE SELECTION (overall usage=85%)
Feature ranking
Filter method
Other FS
Forward / backward wrapper
Embedded method
Wrapper with search
0
10
20
30
40
Percent of participants
50
60
Figure 6: Feature selection methods.
the preferred way of selecting hyper-parameters and performing model selection. But
model selection was to a large extent circumvented by the use of ensemble methods.
Three ensemble methods have been mostly used: boosting (Freund and Schapire,
1996; Friedman, 2000), bagging (Breiman, 1996, 2001), and heterogeneous ensembles
built by forward model selection (Caruana and Niculescu-Mizil, 2004; Caruana et al.,
2006).
Surprisingly, less than 50% of the teams reported using regularization (Vapnik, 1998).
Perhaps this is due to the fact that many ensembles of decision trees do not have explicit
regularizers, the model averaging performing an implicit regularization. The wide majority
of approaches were frequentist (non Bayesian). Little use was made of the unlabeled test
examples for training and no performance gain was reported.
17
18. ´
Guyon, Lemaire, Boulle, Dror, Vogel
CLASSIFIER (overall usage=93%)
Decision tree...
Linear classifier
Non-linear kernel
Other Classif
- About 30% logistic loss,
>15% exp loss, >15% sq loss,
~10% hinge loss.
Neural Network
Naïve Bayes
Nearest neighbors
- Less than 50% regularization
(20% 2-norm, 10% 1-norm).
Bayesian Network
- Only 13% unlabeled data.
Bayesian Neural Network
0
10
20
30
40
Percent of participants
50
60
Figure 7: Classification algorithms.
MODEL SELECTION (overall usage=90%)
10% test
K-fold or leave-one-out
Out-of-bag est
Bootstrap est
Other-MS
- About 75% ensemble methods
(1/3 boosting, 1/3 bagging, 1/3 other).
Other cross-valid
Virtual leave-one-out
- About 10% used unscrambling.
Penalty-based
Bi-level
Bayesian
0
10
20
30
40
Percent of participants
50
60
Figure 8: Model selection methods.
We also analyzed the fact sheets with respect to the software and hardware implementation (Figure 9):
• Hardware: While some teams used heavy computational apparatus, including multiple processors and lots of memory, the majority (including the winners of the slow
track) used only laptops with less than 2 Gbytes of memory, sometimes running in
parallel several models on different machines. Hence, even for the large dataset, it
was possible to provide competitive solutions with inexpensive computer equipment.
In fact, the in-house system of Orange computes its solution in less than three hours
on a laptop.
18
19. Analysis of the KDD Cup 2009
> 8 GB
>= 32 GB
Run in parallel
<= 8 GB
None
<= 2GB
Multi-processor
Memory
Parallelism
Mac
OS
Java
Other (R, SAS)
Matlab
Linux Unix
Windows
C C++
Software Platform
Operating System
Figure 9: Implementation.
• Software: Even though many groups used fast implementations written in C or C++,
packages in Java (Weka) and libraries available in Matlab R or “R”, presumably slow
and memory inefficient, were also widely used. Users reported performing first feature
selection to overcome speed and memory limitations. Windows was the most widely
used operating system, closely followed by Linux and other Unix operating systems.
8. Significance Analysis
One of the aims of the KDD cup 2009 competition was to find whether there are data-mining
methods which are significantly better than others. To this end we performed a significance
analysis on the final results (last submission before the deadline, the one counting towards
the final ranking and the selection of the prize winners) of both the SLOW and FAST track.
Only final results reported on the large dataset were included in the analysis since we have
realized that submissions based on the small dataset were considerably inferior.
To test whether the differences between the teams are statistically significant we followed
a two step analysis that is specifically designed for multiple hypothesis testing when several
independent task are involved (Demˇar, 2006): First we used the Friedman test (Friedman,
s
1937), to examine the null hypothesis H0 , which states that the AUC values of the three
tasks, (Churn, Appetency and Up-Selling) on a specific track (FAST or SLOW) are all
drawn from a single distribution. The Friedman test is a non-parametric test, based on
the average ranking of each team, where AUC values are ranked for each task separately.
A simple test-statistic of the average ranks is sufficient to extract a p-value for H0 ; In the
case when H0 is rejected, we use a two tailed Nemenyi test (Nemenyi, 1963) as a post-hoc
analysis for identifying teams with significantly better or worse performances.
Not surprisingly, if one takes all final submissions, one finds that H0 is rejected with
high certainty (p-value < 10−12 ). Indeed, significant differences are observed even when one
inspects the average final AUCs (see Figure 10), as some submissions were not substantially
19
20. ´
Guyon, Lemaire, Boulle, Dror, Vogel
(a) FAST
(b) SLOW
Figure 10: Sorted final scores: The sorted AUC values on the test set of each of the
three tasks, together with the average of AUC on the three tasks. Only final
submissions are included. (a) FAST track and (b) SLOW track. The baselines for the basic Na¨ Bayes and selective Na¨ Bayes are superposed on the
ıve
ıve
corresponding tasks.
better than random guess, with an AUC near 0.5. Of course, Figure 10 is much less
informative than the significance testing procedure we adopt, which combines the precise
scores on the three tasks, and not each one separately or their averages.
Trying to discriminate among the top performing teams is more subtle. When taking
the best 20 submissions per track (ranked by the best average AUC) - the Friedman test still
rejects H0 with p-values 0.015 and 0.001 for the FAST and SLOW tracks respectively. However, the Nemenyi tests on these reduced data are not able to identify significant differences
between submissions, even with a significance level of α = 0.1!
The fact that one does not see significant differences among the top performing submissions is not so surprising: during the period of the competition more and more teams have
succeeded to cross the baseline, and the best submissions tended to accumulate in the tail
of the distribution (bounded by the optimum) with no significant differences. This explains
why the number of significant differences between the top 20 results decreases with time
and number of submissions.
Even on a task by task basis, Figure 10 reveals that the performance of the top 50%
AUC values lie on an almost horizontal line, indicating there are no significant differences
among these submissions. This is especially marked for the SLOW track.
From an industrial point of view, this result is quite interesting. In an industrial setting
many criteria have to be considered (performance, automation of the data mining process,
training time, deployment time, etc.). But this significance testing shows using state-of the
art techniques, one is unlikely to get significant improvement of performance even at the
expense of a huge deterioration of the other criterions.
20
21. Analysis of the KDD Cup 2009
9. Conclusion
The results of the KDD cup 2009 exceeded our expectations in several ways. First we
reached a very high level of participation: over three times as much as the most popular
KDD cups so far. Second, the participants turned in good results very quickly: within 7
hours of the start of the FAST track challenge. The performances were only marginally
improved in the rest of the challenge, showing the maturity of data mining techniques.
Ensemble of decision trees offer off-the-shelf solutions to problems with large numbers of
samples and attributes, mixed types of variables, and lots of missing values. Ensemble
methods proved to be effective for winning, but single models are still preferred by many
customers. Further work include matching the performances of the top ranking participants
with single classifiers.
Acknowledgments
We are very grateful to the Orange company who donated the data, the computer servers,
many hours of engineer time, and the challenge prizes. We would like to thank all the
Orange team members, including Fabrice Cl´rot and Raphael F´raud. We also gratefully
e
e
acknowledge the ACM SIGKDD and the Pascal2 European network of excellence (FP7-ICT2007-1-216886) who supported the work of the KDD Cup 2009 co-chairs Isabelle Guyon and
David Vogel and the web site development. The support of Google and Health Discovery
Corporation allowed students to attend the workshop. We are very grateful to the technical support of MisterP and Pascal Gouzien. We also thank Gideon Dror for editing the
proceedings.
References
H. Akaike. Information theory and an extension of the maximum likelihood principle. In
B.N. Petrov and F. Csaki, editors, 2nd International Symposium on Information Theory,
pages 267–281. Akademia Kiado, Budapest, 1973.
Bernhard E. Boser, Isabelle Guyon, and Vladimir Vapnik. A training algorithm for optimal
margin classifiers. In COLT, pages 144–152, 1992.
Marc Boull´. Compression-based averaging of Selective Na¨ Bayes classifiers. JMLR, 8:
e
ıve
1659–1685, 2007. ISSN 1533-7928.
Leo Breiman. Bagging predictors. Machine Learning, 24(2):123–140, 1996.
Leo Breiman. Random forests. Machine Learning, 45(1):5–32, 2001.
Rich Caruana, Art Munson, and Alexandru Niculescu-Mizil. Getting the most out of ensemble selection. In Proceedings of the 6th International Conference on Data Mining
(ICDM ‘06), December 2006. Full-length version available as Cornell Technical Report
2006-2045.
21
22. ´
Guyon, Lemaire, Boulle, Dror, Vogel
Rich Caruana and Alexandru Niculescu-Mizil. Ensemble selection from libraries of models.
In Proceedings of the 21st International Conference on Machine Learning (ICML’04),
2004. ISBN 1-58113-838-5.
Clopinet. Challenges in machine learning. URL http://clopinet.cm/challenges.
Janez Demˇar. Statistical comparisons of classifiers over multiple data sets. J. Mach. Learn.
s
Res., 7:1–30, 2006. ISSN 1533-7928.
Yoav Freund and Robert E. Schapire. Experiments with a new boosting algorithm. In
ICML, pages 148–156, 1996.
Jerome H. Friedman. Greedy function approximation: A gradient boosting machine. Annals
of Statistics, 29:1189–1232, 2000.
M. Friedman. The use of ranks to avoid the assumption of normality implicit in the analysis
of variance. Journal of the American Statistical Association, 32:675–701, 1937.
I. Guyon, A. Saffari, G. Dror, and J. Buhmann. Performance prediction challenge. In
IEEE/INNS conference IJCNN 2006, Vancouver, Canada, July 16-21 2006.
T. Hastie, R. Tibshirani, and J. Friedman. The Elements of Statistical Learning, Data
Mining, Inference and Prediction. Springer Verlag, 2000.
IBM Research. Winning the KDD cup orange challenge with ensemble selection. In JMLR
W&CP, volume 7, KDD cup 2009, Paris, 2009.
Hung-Yi Lo et al. An ensemble of three classifiers for KDD cup 2009: Expanded linear
model, heterogeneous boosting, and selective na¨ Bayes. In JMLR W&CP, volume 7,
ıve
KDD cup 2009, Paris, 2009.
Hugh Miller et al. Predicting customer behaviour: The University of Melbourne’s KDD
cup report. In JMLR W&CP, volume 7, KDD cup 2009, Paris, 2009.
T.M. Mitchell. Machine Learning. McGraw-Hill Co., Inc., New York, 1997.
P. B. Nemenyi. Distribution-free multiple comparisons. Doctoral dissertation, Princeton
University, 1963.
V. Vapnik. Statistical Learning Theory. John Wiley and Sons, N.Y., 1998.
Jainjun Xie et al. A combination of boosting and bagging for KDD cup 2009 - fast scoring
on a large database. In JMLR W&CP, volume 7, KDD cup 2009, Paris, 2009.
22